To get a better understanding about how price-setting in the sharing economy works, a wide range of papers have used a hedonic price model to test the consumer valuation of Airbnb listings (e.g. Gibbs et al. (2018), Teubner et al. (2017)). In this kind of modelling, structured attributes (number of rooms, location, rating, etc.) of the listing often together with attributes of the host are used, to evaluate the source of consumer utility.

In the following analysis I want to exploit the textual data in listing description to predict the price of a listing.

Research questions:

  1. Can the use of textmining improve the accuracy of predicting the price of Airbnb listings?
  2. Which keywords affect the price of a listing either positively or negatively?

Method:

To compare my approach with the conventional methods, I first estimate a model in which I use the structured attributes as exogenous regressors to predict the price of an Airbnb listing. Afterwards, I use textual features of the same listings to predict the prices and compare the two models.

The project is divided into three parts. In this section I describe the data set and how I prepare it for analysis. In the second part I estimate a linear model with the conventional attributes and in the third part I use text data for the same listings.

Load Data

I use a unique dataset that contains information on 47.006 Airbnb listings from seven major German cities, namely Berlin, Munich, Hamburg, Cologne, Dresden, Stuttgart and Frankfurt am Main. Listings were gathered directly from Airbnb’s website in September 2017 using a custom web scraper. In this way I have obtained all publicly available information for a listing, including but not limited to prices, accommodation features, reviews and host details.

Data Preparations

head(rooms)
## # A tibble: 6 x 62
##    room_id  host_id room_type  country city  neighborhood address    price
##      <int>    <int> <chr>      <chr>   <chr> <chr>        <chr>      <int>
## 1 19117409   1.34e⁸ Entire ho… Deutsc… Hamb… <NA>         Othmarsch…   129
## 2  5728058   3.34e⁵ Entire ho… Deutsc… Hamb… <NA>         Neustadt,…   116
## 3 19954984   1.41e⁸ Entire ho… Deutsc… Münc… <NA>         Schwabing…    91
## 4  9918551   5.10e⁷ Entire ho… Deutsc… Schö… <NA>         Schönefeld    43
## 5 13836114   8.16e⁷ Entire ho… Deutsc… Hamb… <NA>         Eimsbütte…    61
## 6 20355318   8.02e⁷ Entire ho… Deutsc… Köln  <NA>         Köln          49
## # ... with 54 more variables: nightly_price <int>, reviews <int>,
## #   accommodates <int>, bathrooms <int>, bedrooms <int>, bed_type <chr>,
## #   minstay <int>, last_modified <dttm>, latitude <dbl>, longitude <dbl>,
## #   survey_id <int>, location <chr>, coworker_hosted <chr>,
## #   extra_host_languages <chr>, name <chr>, property_type <chr>,
## #   currency <chr>, rate_type <chr>, overall_satisfaction <chr>,
## #   cleanliness_satisfaction <int>, communication_satisfaction <int>,
## #   location_satisfaction <int>, accuracy_satisfaction <int>,
## #   checkin_satisfaction <int>, value_satisfaction <chr>, amenities <chr>,
## #   cancel_policy <chr>, instant_book <chr>, response_time <chr>,
## #   response_rate <dbl>, friend_count <int>, wishlist_count <int>,
## #   pic_count <chr>, superhost <chr>, description_language <chr>,
## #   hostname <chr>, rule_children <chr>, rule_infants <chr>,
## #   rule_pets <chr>, rule_smoking <chr>, rule_events <chr>,
## #   hostprofilepic <chr>, cleaning_fee <chr>, security_deposit <chr>,
## #   last_review <dttm>, positive_reviews <dttm>, negative_reviews <date>,
## #   last_cal_update <chr>, member_since <chr>, host_verified <chr>,
## #   deleted <chr>, filled <chr>, description <chr>, base_price <chr>
# Convert strings to numeric
rooms <- rooms %>% 
  mutate(overall_satisfaction = as.numeric(overall_satisfaction),
         pic_count = as.numeric(pic_count)) %>%
  filter(!is.na(overall_satisfaction))

(1) Cities

Keep only listings from the following cities: Hamburg, München, hamburg, Köln, FFM, Dresden, Stuttgart

## create clean-up function
create_city <- function(x, city){
  city_clean <- ifelse(grepl(x, city),x , city) 
  return(city_clean)
}
city_list <- c("Hamburg","München","Berlin","Frankfurt","Köln","Stuttgart","Dresden")

for(i in city_list){
  rooms$city <- create_city(i, rooms$city)
}

rooms %>%
  filter(city %in% city_list) -> rooms

rooms %>%
  group_by(city) %>%
  tally() %>%
  ggplot(aes(reorder(city, n, desc),n)) +
  geom_col(fill = col[3], alpha = 0.8) +
  labs(x="", y="", title="Count")

(2) Property Type

rooms %>%
  group_by(property_type) %>%
  tally() %>%
  ggplot(aes(reorder(property_type, n),n)) +
  geom_col(fill = col[3], alpha = 0.8) +
  labs(x="", y="", title="Property Types") +
  coord_flip()

To keep things simple, I will just keep listings of property type “Wohnung” (apartment)

rooms %>%
  filter(property_type == "Wohnung") -> rooms

(3) Roomtype

rooms %>%
  ggplot(aes(room_type)) +
  geom_bar(fill = col[3], alpha = 0.8) +
  labs(x="", y="")

(4) Price

rooms %>%
  ggplot(aes(city, price)) +
  geom_boxplot(outlier.size = 0)

Apparently, there are some outliers. After cheking the respective listings, I decided to exclude them.

rooms %>%
  filter(price < 1500) -> rooms
rooms$price.cut <- cut(rooms$price, c(seq(0,500,1), Inf))

rooms %>%
  ggplot(aes(as.numeric(price.cut), factor(city))) +
  geom_density_ridges(scale = 5,
                      fill = col[3], alpha = 0.7,
                      color = "white") +
  theme_ridges() +
  scale_x_continuous(expand = c(0, 0), labels = c(seq(0,400,100),">500")) +
  labs(y="", x="Price")

(5) Rating

rooms %>%
  ggplot(aes(overall_satisfaction, factor(room_type))) +
  geom_density_ridges(scale = 5,
                      fill = col[3], alpha = 0.7,
                      color = "white") +
  scale_x_continuous(expand = c(0, 0)) +
  labs(y="", x="Rating")

(6) Number of Reviews

Next, I exclude listings with less than three reviews, as it can be assumed that these listings have never been booked, or only very little.

rooms %>% 
  filter(reviews >= 3) -> rooms
rooms$reviews.cut <- cut(rooms$reviews, c(seq(0,50,1), Inf))

rooms %>%
  ggplot(aes(as.numeric(reviews.cut), factor(city))) +
  geom_density_ridges(scale = 5,
                      fill = col[3], alpha = 0.7,
                      color = "white") +
  scale_y_discrete(expand = c(0,0)) +
  scale_x_continuous(expand = c(0,0),
                     breaks = c(seq(0,50,10)),
                     labels = c(seq(0,40,10),">50")) +
  labs(y="", x="Number of Reviews")

Final dataframe

df <- rooms %>% 
  select(room_id, name, 
         description, city, price, overall_satisfaction,
         room_type, bed_type, pic_count,
         reviews, accommodates, bedrooms, minstay,
         latitude, longitude) %>%
  mutate(fulltext = paste(name, description, sep=" "))

Textdata

Turning to the text data, lets first have a quick look at three random descriptions:

rooms %>% sample_n(3) %>%
  select(description) %>%
  knitr::kable(align = "l")
description
Mein Zimmer in einer WG ist vom 18.04-22.04 frei. Darum vermiete ich dieses für den Zeitraum unter. Ich wohne mit 3 sehr freundlichen Mitbewohnern (22-25). Sie sprechen sprechen alle fließend Deutsch und Englisch.
Hallo! Willkommen in meiner sonnigen, ruhigen und gemütlichen 2-Zimmer Wohnung in Wilmersdorf! Wer wie Zuhause wohnen will, ist hier richtig!
Beautiful modern design apartment in the heart of Munich - fully equipped kitchen - spacious living room with balcony facing the courtyard. Perfect location - 2 minutes - to the famous Maximilianstrasse - Underground parking available upon request. Schöne moderne Design- Wohnung im Herzen von München - voll ausgestattete Küche - großzügiges Wohnzimmer mit Balkon zum Innenhof. Perfekte Lage in der Nähe - 2min. - zur berühmten Maximilianstrasse - Tiefgaragenstellplatz verfügbar.

Languages

In which languages are the descriptions written?

load(file = "../output/prep1.Rda")
df %>% group_by(language) %>% 
  tally() %>%
  ggplot(aes(reorder(language, n),n)) +
  geom_col(fill = col[3], alpha = 0.7) +
  coord_flip() +
  labs(x="",y="")

Check sample articles if the classification is valid

df %>%
  sample_n(5) %>%
  select(fulltext, language) %>%
  knitr::kable()
fulltext language
Helle Wohnung gepflegt und ruhig Schöne, neu eingerichtete helle 1 Zimmer Wohnung mit komplett neuem Interieur und Elektrogeräten. Das Badezimmer ist weiss gefliest, mit Fenster. Innerhalb weniger Gehminuten erreichst Du alle Restaurants und Cafes in Friedrichshain. german
Bright room in 1870s-building, nice neighbourhood In einem Altbau von 1890 seid ihr zu Gast in einem großen, hellen Zimmer. Schön eingerichtet für jede und für jeden, der Farben, Bücher, Musik und Katzen mag. german
Zentral, ruhig, hell, viel Platz In der 1. Etage eines Einfamilienhauses befindet sich eine eigene Wohnung mit 2 Schlafräumen. Ab 3 Personen stehen 3 Schlafräume zur Verfügung (20, 22 bzw. 25 qm) . Dazu gibt es eine Küche , in der gekocht und gegessen werden kann , ein Bad mit Dusche und Wanne und eine Terrasse . Das Haus ist von einem Garten umgeben und befindet sich im grünen Stadtteil Dresden- Strehlen . Um das Stadtzentrum von Dresden zu erreichen , stehen eine Bus- und 2 Straßenbahnlinien zur Verfügung (Haltestelle in 3 Minuten und Fahrtzeit 10-15 Minuten). Der nahe Bahnhof von Dresden - Strehlen lädt zu Fahrten nach Meißen und in die Sächs. Schweiz ein. Etliche Gaststätten , 1 Supermarkt (Konsum) und eine gute Bäckerei sorgen für den täglichen Bedarf. Ich bewohne das Erdgeschoß und vermiete die 1. Etage beinahe ganzjährig als Ferienwohnung. Ich bin Urdresdner und zeige meinen Gästen gern bei Bedarf die Stadt und ihre schöne Umgebung. Seit Februar 2017 erhebt die Stadt Dresden eine Beherbergungssteuer, in meinem Fall 1-2 Euro pro Person und Nacht. Das Geld ist bei Anreise gegen Quittung an mich bar zu zahlen. Ich gebe es dann an die Stadt weiter. Zeit für An- und Abreise kann bei Bedarf individuell geregelt werden. Während des Aufenthalts der Gäste stehe ich gern als Ansprechpartner zur Verfügung. german
Loft with Huge Terrace This apartment is simply the best: Top location, great design, nice kitchen (all appliances incl. dish washer & washing machine), wonderful king size bed, bathroom with tub. The terrace offers a lovely view into Berlins first bycicle-street. english
1 bedroom Prenzlauer Berg Cosy and bright City Apartment in the hart of Prenzlauer Berg, closed to Schönhauser Allee. Spacios 50m2 with 2 rooms to chare , large kittchen. There is a 2nd bed/couch. You dont need a car, everything you find in walking distance, Bars & Restaurants, train Station Schönhauser Allee, Cinemax, Shopping Center. english

Ok, looks good. Lets only keep listings with german and english descriptions.

df %>%
  filter(language %in% c("german","english")) -> df
ggplot(df, aes(x=factor(city))) +
  geom_bar(aes(fill = language),
           alpha = 0.8) +
  labs(x="", y="", fill="")

It is not surprising that Berlin seems to be the most international city, measured by the listings that have their description in English. But I am a little disappointed with Hamburg…

Word count

How long are the descriptions on average?

df$text_length <- sapply(gregexpr("\\S+", df$fulltext), length)
df$text_length.cut <- cut(df$text_length, c(seq(0,150,1),Inf))

df %>%
  ggplot(aes(as.numeric(text_length.cut), factor(city))) +
  geom_density_ridges(aes(fill = language),
                      color = "white", alpha = 0.8) +
  scale_x_continuous(expand = c(0,0), 
                     labels = c(seq(0,100,50),">150")) +
  labs(y = "", x = "Word Count", fill= "") +
  theme()

Surprisingly, the English texts are longer.

Pre-Processsing

Next, I have to pre-process the text data to be able to include it into my model. Text data is inherently high-dimensional, so to reduce this dimensionality the following steps will be applied:

  1. Remove Punctuation, Numbers,…
  2. Stopword removal: Stopwords (highly frequent terms like “and”, “or”, “the”) are stripped out of text as they do add any helpfull information about the listing.
  3. Tokenization: splitting of a raw character string into individual elements of interest: words, numbers, punctuation.
  4. Document Term Matrix Represent each listing as a numerical array of unique terms (bag-of-words model). This will be done in part three of this project.

(1) Remove Punctuation, Numbers, …

df$text_cleaned <- gsub("[[:punct:]]", " ", df$fulltext)
df$text_cleaned <- gsub("[[:cntrl:]]", " ", df$text_cleaned)
df$text_cleaned <- gsub("[[:digit:]]", " ", df$text_cleaned)
df$text_cleaned <- gsub("^[[:space:]]+", " ", df$text_cleaned)
df$text_cleaned <- gsub("[[:space:]]+$", " ", df$text_cleaned)
df$text_cleaned <- tolower(df$text_cleaned)

(2) Remove Stopwords

df$text_cleaned <- removeWords(df$text_cleaned, stopwords("english"))
df$text_cleaned <- removeWords(df$text_cleaned, stopwords("german"))

(3) Tokenizing

Unigrams

token.df <- df %>%
  tidytext::unnest_tokens(word, text_cleaned) %>%
  filter(nchar(word) > 1) %>%
  filter(nchar(word) < 30)

token.df %>% 
  count(word, sort = TRUE) %>%
  ungroup() %>%
  top_n(20, n) %>%
  knitr::kable(align="l")
word n
wohnung 12264
apartment 9732
zimmer 8800
room 8529
min 8365
berlin 5994
bahn 5187
restaurants 4511
minuten 4289
flat 4200
küche 3877
city 3862
nähe 3800
unterkunft 3488
bars 3228
qm 3060
direkt 2992
liegt 2983
station 2955
lage 2916

Bigrams

bigram.df <- df %>%
  unnest_tokens(bigram, text_cleaned, 
                          token = "ngrams", n=2) 

bigram.df %>% 
  count(bigram, sort = TRUE) %>%
  ungroup() %>%
  top_n(20, n) %>%
  knitr::kable(align="l")
bigram n
u bahn 2699
s bahn 1870
zimmer wohnung 1497
wohnung liegt 1287
prenzlauer berg 1083
living room 1081
city center 989
walking distance 982
unterkunft gut 936
bars restaurants 891
paare alleinreisende 848
gut paare 832
unterkunft nähe 811
restaurants bars 786
alleinreisende abenteurer 771
wohnung befindet 751
unmittelbarer nähe 745
unterkunft lieben 733
st pauli 689
lieben wegen 678

Wordclouds

corp <- corpus(df$text_cleaned)
docvars(corp)<-df$city   #attaching the class labels to the corpus message text

col <- RColorBrewer::brewer.pal(10, "BrBG")  

(1) Berlin

c.plot <- corpus_subset(corp, docvar1=="Berlin")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 250, color = col)

(2) Hamburg

c.plot <- corpus_subset(corp, docvar1=="Hamburg")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 200, color = col)

(3) München

c.plot <- corpus_subset(corp, docvar1=="München")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 50, color = col)

(4) Köln

c.plot <- corpus_subset(corp, docvar1=="Köln")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 50, color = col)

(5) Frankfurt

c.plot <- corpus_subset(corp, docvar1=="Frankfurt")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 50, color = col)

(6) Stuttgart

c.plot <- corpus_subset(corp, docvar1=="Stuttgart")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 50, color = col)

(7) Dresden

c.plot <- corpus_subset(corp, docvar1=="Dresden")
c.plot<-dfm(c.plot, tolower = TRUE, remove_numbers = TRUE, remove=stopwords("SMART"))

textplot_wordcloud(c.plot, min.freq = 50, color = col)